Dropping an identity document's name vectors scans every fact's evidence #1408

name_vectors::drop_for_document (#1390) deletes the vectors of facts with evidence from one document, and the flag flip in sources.rs does the same for a source. fact_evidence's only index is its primary key (fact_id, chunk_id) (migrations/0003_graph.sql), so both scan the whole table. A code mirror that pushes thousands of identity documents does that once per push. Worth measuring on a large base the way #1296 was, and adding an index on document_id if it shows.